Tools and setup
Uses
A short list, sorted by how much evidence sits behind each entry. Production use comes first, then personal projects, then what I'm exploring. A tool listed here is not a claim of expertise unless it is in the first group.
Last updated
Used extensively in production
Years of use in shipped systems at Farfetch, Wellhub, and earlier roles.
Modelling
PyTorch Sequence models with attention for Tailor at Farfetch; also my research work.
scikit-learn · XGBoost · LightGBM Classical and tabular ML: churn, segmentation, and baselines.
Python scientific stack NumPy, pandas, and SciPy: analysis, Bayesian models, and statistics.
Experimentation & evaluation
A/B testing & causal impact Validated size-and-fit features at Farfetch, including the Fit Advice return reductions.
Contextual bandits Theory and production considerations in the published series.
Opik · DeepEval · RAGAS Tools for evaluating LLM and RAG applications, including LLM-as-a-judge.
MLflow Experiment tracking and model registry.
Data & infrastructure
Spark / PySpark · Databricks Training data and ETL pipelines at Farfetch and Sonae.
SQL / BigQuery Analytics and feature queries.
Airflow · Kubeflow Training and data pipeline orchestration.
Kafka Real-time event processing.
Vespa.ai Real-time serving and ranking for recommendations at Farfetch.
Azure · AWS · Docker Cloud and container tooling across professional roles.
Used in personal projects
Public work on this site, but no production use behind it.
uv · ruff · pytest Tooling for the verifiable-rewards study and its 112-test transformer.
einops Readable tensor reshapes; see the einops note in the Notebook.
LlamaIndex · Elasticsearch · Ollama A local-first RAG prototype.
MLX An on-device chat experiment on Apple Silicon.
Currently exploring
I've read the code or docs of these for my writing, but nothing shipped or published rests on them yet.
TRL · verl Reference GRPO/PPO implementations to check my own against.
Vowpal Wabbit Production bandit library, studied for the contextual bandits series.